Wrist joint bone classification method based on deep learning

Abstract

The invention relates to the technical field of deep learning, solves the technical problem that when an existing deep learning method is used for processing a medical image, refinement of data preprocessing and universality of a training strategy are often ignored, so that the generalization ability of a model is limited, and particularly relates to a wrist joint bone classification method based on deep learning. Firstly, a standardized preprocessing process is performed on a medical image, and secondly, a two-stage hierarchical data enhancement mechanism is introduced to greatly improve the model performance. Then, a VGG16bn model based on ImageNet pre-training is selected and finely adjusted to serve as a core feature extractor, and a Focal Loss loss function is introduced in the model training process, so that key features of osteoporosis in a medical image can be accurately captured; finally, in the verification of an osteoporosis image classification task, the accuracy rate in the classification of wrist joints is as high as 98%, and excellent performance and wide application prospects in clinical diagnosis are manifested.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a deep learning-based method for classifying wrist joint bones. Background Technology

[0002] In recent years, the rapid development of deep learning technology has gradually revealed its enormous potential in many fields, especially in key applications such as autonomous driving, facial recognition, and medical image classification. The integration of deep learning with these technologies has become an important way to improve efficiency and accuracy. Taking medical image classification as an example, traditionally, doctors rely on extensive professional knowledge and clinical experience to diagnose illnesses. This process not only severely tests the doctors' abilities but also cannot completely avoid the risk of misdiagnosis.

[0003] The intervention of deep learning, through systematic learning and training on massive amounts of real medical image data, can build high-precision prediction models, thereby significantly reducing the misdiagnosis rate of diseases and effectively improving doctors' diagnostic efficiency. Furthermore, the prediction results provided by deep learning models trained on large-scale data, due to their strong data support and objective analysis capabilities, can also bring patients greater trust and fewer doubts.

[0004] The deep integration of deep learning with the medical field is both an inevitable trend of the times and an urgent need to be addressed in the current healthcare sector. However, the inherent high complexity of medical images and the inconsistent image quality have always been key bottlenecks hindering the effective realization of this process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for wrist joint bone classification. This method solves the technical problem that existing deep learning methods often neglect the refinement of data preprocessing and the universality of training strategies when processing medical images, resulting in limited model generalization ability.

[0006] To address the aforementioned technical challenges, this invention proposes a deep learning-based method for wrist joint bone classification. This method introduces a customized multi-stage data preparation process, combined with the classic and efficient VGG16_bn convolutional neural network architecture, significantly improving the classification model's ability to identify subtle features of osteoporosis and enhancing the accuracy of clinical diagnosis. In the preprocessing stage, not only are medical images standardized in format, window width and level normalized, and size unified, but a two-stage data augmentation strategy is also innovatively implemented. This greatly enriches the diversity of training samples, effectively suppressing model overfitting. For the potential class imbalance problem in medical image data, the Focal Loss loss function is used to strengthen the learning of easily confused categories, significantly improving the generalization performance of the classification model.

[0007] To achieve the above-mentioned technical objectives, the present invention adopts the following steps: Step 1: Standardized Preprocessing of Medical Images. This step aims to convert the original heterogeneous medical images or videos into a unified, high-quality digital image format. The specific process is as follows: First, the DICOM format medical image files are precisely parsed using the pydicom tool to comprehensively extract their pixel data and associated metadata. Second, window width and window level normalization operations are performed on the extracted pixel data, linearly mapping the original high dynamic range pixel values ​​to the standard 0-255 range. The processing logic for MultiValue pixels is specifically optimized to ensure that each pixel is calculated with a single, accurate value. Finally, the processed medical images are scaled proportionally to a target resolution of 1024×1024, and saved as .png files while maintaining the original image proportions and avoiding distortion.

[0008] Step 2: Hierarchical Dynamic Data Augmentation and Dataset Construction. This step aims to significantly improve the robustness and generalization ability of the classification model through multi-dimensional augmentation strategies, while optimizing the medical image dataset.

[0009] 2.1 Balanced partitioning of medical image dataset: First, the standardized medical image dataset is precisely partitioned according to image categories (normal bone mass, osteoporosis, and osteopenia). The ratio of training set to validation set is set to 85%:15%, and the balanced distribution of samples of each category in the original data is strictly ensured in each set to effectively alleviate the common class imbalance problem in deep learning.

[0010] 2.2 Unified Preset Enhancement: The first stage of unified preset enhancement is implemented on the segmented medical image dataset. This stage integrates adaptive local contrast enhancement, edge feature enhancement, and pseudo-color rendering, including: contrast-limited adaptive histogram equalization to enhance local contrast; edge feature enhancement to highlight bone contours; pseudo-color rendering to visualize bone density differences through color mapping; and brightness and contrast adjustment. These enhancement operations aim to enrich image features globally and compensate for uneven illumination and low contrast issues during image acquisition.

[0011] 2.3 Real-time Random Augmentation: In each epoch of VGG16_bn model training, a second stage of real-time random data augmentation is dynamically applied to the medical images in the training set. This includes random horizontal flipping, random vertical flipping, random rotation within the range of [-20°, +20°], and random cropping and scaling. This strategy ensures that the classification model is exposed to unique image variants in each iteration, further simulating the diversity of images in the real world, thereby significantly enhancing the classification model's robustness and generalization ability.

[0012] Step 3: Training the intelligent model for osteoporosis classification based on the VGG16_bn model. This step focuses on building and optimizing the core model.

[0013] 3.1 Model Initialization and Input Adaptation: A VGG16_bn model pre-trained on the ImageNet large-scale dataset was loaded as the backbone network for deep feature extraction. This pre-trained weight endowed the model with powerful general visual feature recognition capabilities. Simultaneously, images that underwent two-stage data augmentation were uniformly adjusted to the standard input size of 224×224 required by the VGG16_bn model.

[0014] 3.2 Fine-tuning Training and Loss Function Optimization: The VGG16_bn model was fine-tuned using the training set data processed in step 2 to ensure its feature extraction capabilities accurately match the characteristics of osteoporosis medical images. During training, to effectively address the potential class imbalance problem in the medical image dataset, a Focal Loss loss function was innovatively introduced. Focal Loss dynamically adjusts sample weights, allowing the VGG16_bn model to focus more on the minority class samples that are difficult to classify, thereby improving the classification accuracy of the model for all classes, especially the minority classes. The VGG16_bn model predicts the probability of each class through forward inference of the neural network and continuously optimizes the network parameters through backpropagation.

[0015] Step 4: Model Validation and Performance Evaluation. This step aims to objectively evaluate the actual performance of the trained model. The validation set data from the medical image dataset processed in Steps 1 and 2 is loaded and input into the trained classification model. The classification model calculates the probability of each image belonging to each category through forward inference and the SoftMax function. Then, it performs a weighted summation based on uniformly enhanced weights, selecting the category with the highest probability as the final classification prediction. By comparing the model's classification results with the true labels and conducting statistical analysis in multiple rounds of experiments, the classification accuracy of the model is finally obtained, quantifying its effectiveness and potential in osteoporosis diagnosis.

[0016] By employing the above technical solution, the present invention provides a wrist joint bone classification method based on deep learning, which has at least the following beneficial effects: First, this invention significantly improves the quality and diversity of medical image data through a rigorous image standardization preprocessing and an innovative two-stage hierarchical data augmentation strategy. Specifically, the preprocessing stage ensures the uniformity of image format and the accuracy of pixel values, while the uniquely designed two-stage data augmentation greatly enriches the representation of training samples, effectively suppresses the risk of overfitting during model training, and significantly enhances the model's generalization ability and robustness when faced with unknown and diverse medical images. This not only alleviates the inherent complexity and inconsistent quality of medical image data but also provides a solid foundation for the model to learn the subtle features of osteoporosis.

[0017] Secondly, this invention achieves high-precision and stable osteoporosis classification by integrating the efficient VGG16_bn model and combining it with Focal Loss optimization. Utilizing the VGG16_bn model pre-trained on ImageNet, its powerful general feature extraction capabilities significantly shorten the training time from scratch, greatly improving training efficiency. More importantly, addressing the common class imbalance problem in medical image classification tasks, it innovatively introduces the Focal Loss loss function, enabling the classification model to more effectively focus on and learn from the few difficult-to-distinguish minority class samples. This ensures high overall accuracy while enhancing the classification model's comprehensive ability to identify various degrees of osteoporosis.

[0018] Finally, the application of this invention will bring significant clinical and social benefits. Automated, high-precision deep learning classification models can significantly reduce the diagnostic burden on doctors, improve their work efficiency, and substantially reduce the potential risk of misjudgment during manual diagnosis. If deployed in medical institutions, this invention will allow doctors to focus their valuable energy on handling more urgent and complex cases, optimize the allocation of medical resources, and ultimately improve the overall quality of healthcare services. This objective and authoritative AI-assisted diagnostic result also helps enhance patients' trust in the diagnostic results, reduce doubts, and thus improve doctor-patient communication, possessing profound clinical application value and positive social significance. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the deep learning-based wrist joint bone classification method of this invention. Figure 2This is a schematic diagram of the wrist joint after using unified data augmentation according to the present invention; Figure 3 This is a schematic diagram of the wrist joint after augmentation using random data according to the present invention; Figure 4 This is a diagram showing the results of fine-tuning the VGG16_bn model in a medical image dataset according to the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0021] Given the challenges of current medical image classification tasks, traditional osteoporosis diagnosis typically relies on high-tech equipment such as dual-energy X-ray absorptiometry (DXA) to measure bone density and make accurate judgments based on T-score values. However, such equipment is generally expensive and difficult to deploy in rural or remote areas, limiting its widespread application. Furthermore, relying solely on physicians' experience and expertise makes it difficult to achieve consistently high accuracy when dealing with subtle changes in bone density. To address these practical challenges, this invention proposes a deep learning-based method for wrist joint bone classification. Figure 1 As shown, this method achieves high-accuracy classification of osteoporosis by performing refined preprocessing on medical images, combining a unique multi-stage data augmentation strategy, and optimizing with an efficient VGG16_bn model and Focal Loss loss function.

[0022] Example 1 This embodiment aims to provide a low-cost, high-efficiency, and highly reliable auxiliary diagnostic tool. Through innovative optimizations and in-depth improvements in both image preprocessing and model training stages, an accuracy rate of up to 98% was achieved in the wrist osteoporosis image classification task, fully demonstrating the effectiveness and clinical application potential of the proposed method.

[0023] The method flow of this embodiment mainly covers three core stages: data preparation and enhancement, model construction and training, and model validation and evaluation. In the data preparation and enhancement stage, the original DICOM format medical images are first standardized, including using pydicom to parse and obtain pixel data and metadata, normalizing window width and window level (handling MultiValue cases), scaling the medical images proportionally to 1024×1024 and saving them as .png format. Subsequently, two stages of data enhancement are implemented: the first stage performs unified preset enhancements such as adaptive local contrast enhancement, edge feature enhancement, and pseudo-color rendering, such as... Figure 2As shown. In the second stage, during each epoch of the VGG16_bn model training, real-time random augmentation operations such as random horizontal flipping, random vertical flipping, random rotation within the range of [-20°, +20°], and random cropping and scaling are dynamically injected, as shown. Figure 3 As shown in Table 1, these enhancements collectively construct a diverse medical image dataset. The training set used in this embodiment comprises 2517 images, covering three categories: normal bone mass, reduced bone mass, and osteoporosis. The osteoporosis sample size is significantly larger than the other categories, exhibiting a clear data imbalance. In this embodiment, the constructed medical image dataset is divided into an 85% training set and a 15% validation set to ensure a balanced distribution of samples across the training and validation sets. For example, it can be divided into approximately 2138 training images and 379 validation images.

[0024] Table 1 shows the accuracy of classification for different categories in the training set. In the model building and training phase, this embodiment uses the VGG16_bn model pre-trained on the ImageNet dataset as the backbone network. This model is fine-tuned to adapt to the feature distribution of medical images, and the input images are uniformly adjusted to a size of 224×224. To address the significant class imbalance problem in the training data, this embodiment innovatively introduces the Focal Loss function. By dynamically adjusting the sample weights, the VGG16_bn model training process focuses more on the few difficult-to-classify classes, thereby improving the recognition accuracy for all classes. Figure 4 As shown, the entire training process uses a portion of the training set as input to the neural network. Forward propagation calculates the classification probabilities of various skeletal health conditions (normal bone mass, osteoporosis, and reduced bone mass), and backpropagation is performed using the Focal Loss function to continuously optimize the network parameters. The use of a pre-trained model significantly reduces training time and computational resource consumption in the initial stages of the model development.

[0025] In the model validation and evaluation phase, the processed validation set is loaded into the trained classification model. The output of the classification model is normalized into probabilities for each class using the SoftMax function, and the class with the highest probability is selected as the classification result. Finally, by comparing the predicted results with the true labels, the final classification accuracy of the classification model is statistically analyzed in multiple rounds of experiments to quantify its actual effectiveness in osteoporosis diagnosis. Through the synergistic effect of the above three stages, this embodiment successfully constructs a neural network model with high accuracy, high robustness, and suitability for routine hospital work in osteoporosis classification tasks.

[0026] Based on the above, the specific steps for training the VGG16_bn model are as follows: Step 1: Standardized preprocessing and format conversion of medical images.

[0027] Furthermore, the original format of medical images is DICOM (.DCM), which is the industry standard for medical image storage and transmission. To facilitate subsequent image processing and deep learning model training, this embodiment first uses the pydicom tool to parse the .DCM file, comprehensively extracting the pixel data and associated metadata of the image. Subsequently, the original pixel values ​​are normalized by window width and window level, linearly mapping them to a display range of 0-255, and special handling is performed on MultiValue pixels to ensure that each pixel has a single and accurate intensity value. This is crucial for adjusting the contrast and brightness of medical images. Next, the normalized image is proportionally scaled to a target size of 1024×1024 pixels to ensure image integrity and avoid feature loss due to deformation. Finally, the processed images are uniformly saved in .png format for easy reading and manipulation by general image processing libraries.

[0028] Step 2: Multi-stage data augmentation.

[0029] Two-stage data augmentation operations are performed on preprocessed medical images to construct an efficient and robust medical image dataset.

[0030] 2.1 Unified Preset Enhancement: The first stage of preset enhancement processing is performed on the preprocessed medical images. This stage integrates adaptive local contrast enhancement, edge feature enhancement, and pseudo-color rendering enhancement to enrich the image feature representation from different dimensions and reduce the model prediction instability caused by differences in image acquisition, such as... Figure 2 As shown.

[0031] 2.2 Real-time Random Augmentation and Dataset Partitioning: For medical images augmented with a uniform preset, a second stage of real-time random data augmentation is dynamically performed in each epoch of the VGG16_bn model training. This includes random horizontal and vertical flipping, random rotation within the range of [-20°, +20°], and random cropping and scaling, such as... Figure 3 As shown in Table 1, this strategy greatly increases the diversity and complexity of the training samples. Finally, based on the characteristics of the training set used in this embodiment, all processed medical images are constructed into a medical image dataset, and divided into an 85% training set and a 15% validation set to ensure a balanced distribution of each category (normal bone mass, reduced bone mass, osteoporosis) in both the training and validation sets. For example, a total of 2517 images can be divided into approximately 2138 training images and 379 validation images to ensure the diversity of training and the stability of validation.

[0032] Step 3: Osteoporosis classification training and optimization based on the VGG16_bn model.

[0033] Furthermore, this step aims to build and train a classification model for intelligent osteoporosis classification.

[0034] 3.1 Model Loading and Structure Adaptation: The VGG16_bn model, pre-trained on the ImageNet large-scale dataset, is loaded as the backbone network. The VGG16_bn model, with its deep convolutional architecture and Batch Normalization layers, possesses powerful general visual feature extraction capabilities and training stability. To adapt to this medical image three-class classification task (normal bone mass, osteoporosis, and osteopenia), the top-level classifier of the VGG16_bn model is modified into a fully connected layer that outputs three classes. The medical images enhanced in step 2 will be resized to 224×224 pixels to fit the standard input size of the VGG16_bn model.

[0035] 3.2 Efficient Training and Focal Loss Optimization: A transfer learning strategy was adopted, using the training set constructed in step 2 to fine-tune the VGG16_bn model. This process utilizes pre-trained weights as model initialization parameters, significantly reducing the time and computational resources required for training from scratch. During the backpropagation phase, this practical example innovatively introduces the FocalLoss loss function to evaluate the difference between the classification model output and the true label, and uses the SGD optimizer to dynamically adjust the network weights. The FocalLoss loss function, through dynamic weighting, effectively solves the common class imbalance problem in medical image datasets. For example, in Table 1, the number of osteoporosis samples far exceeds that of other categories, enabling the VGG16_bn model to pay more attention to the few difficult-to-classify samples during training, thus achieving more balanced and high-precision recognition across all categories. The VGG16_bn model calculates the probability of belonging to three skeletal health conditions through forward inference on the input medical image and updates the network parameters accordingly, gradually improving the prediction accuracy. Figure 4 As shown.

[0036] Step 4: Model performance verification and evaluation of the VGG16_bn model.

[0037] Furthermore, this step aims to objectively and systematically evaluate the actual classification ability of the classification model.

[0038] 4.1 Validation Set Inference and Probability Output: The validation set constructed in step 2 is loaded into the trained classification model. The classification model performs forward inference on each medical image in the validation set and transforms the original score vector into a normalized probability distribution through the SoftMax function connected to the output layer. This function generates a probability value between 0 and 1 for each category, and the sum of the probabilities of all categories is 1, intuitively representing the model's confidence that the image belongs to a specific category. Finally, the classification model selects the category with the highest probability value as the predicted classification result for that image.

[0039] 4.2 Accuracy Statistics and Comprehensive Evaluation: By comparing the predicted classification results of the classification model with the true labels of the images in the validation set one by one, the number of correctly classified samples is recorded, and the classification accuracy of the model is calculated. To ensure the stability and reliability of the evaluation results, this embodiment statistically analyzes and averages the accuracy in multiple rounds of trials to reduce errors caused by randomness. Furthermore, other evaluation metrics, such as precision, recall, F1 score, ROC curve, and AUC value, can be combined to conduct a comprehensive and in-depth performance analysis of the classification model. These metrics are particularly relevant for imbalanced datasets, as they more accurately reflect the model's ability to identify samples from each class. These comprehensive evaluation results provide a solid experimental basis and theoretical support for the practical application of the model in clinical auxiliary diagnosis.

[0040] Example 2 This embodiment aims to verify the effectiveness and superiority of the intelligent osteoporosis classification method proposed in this invention. The experimental data comes from a self-made medical image dataset, which contains approximately 2517 wrist joint images in raw DICOM format. The specific distribution is shown in Table 1 of Embodiment 1, and the images are labeled as three categories: normal bone mass, reduced bone mass, and osteoporosis.

[0041] In the data preparation stage, this embodiment first performs standardized preprocessing on the original DICOM format medical images, including pydicom parsing, window width and window level normalization (handling MultiValue cases), and proportional scaling to 1024×1024 pixels and saving as a .png file. This invention implements a two-stage data augmentation strategy: first, it performs unified preset enhancements such as adaptive local contrast enhancement, edge feature enhancement, and pseudo-color rendering, as well as... Figure 2 As shown. These methods work together to not only highlight image features but also significantly expand the dataset. Subsequently, in each epoch of the VGG16_bn model training, real-time random augmentation operations such as random horizontal flips, random vertical flips, random rotations within the range of [-20°, +20°], and random cropping and scaling are dynamically injected, further enriching the diversity of the training samples, such as... Figure 3 As shown. Finally, the preprocessed and two-stage augmented images were constructed into a medical image dataset, divided into a training set of approximately 2138 images and a validation set of approximately 379 images, for use in training and performance evaluation of the VGG16_bn model, respectively.

[0042] In the model building and training phases, this embodiment uses the VGG16_bn model pre-trained on the ImageNet large-scale dataset as the core classifier. This model, with its superior performance on general vision tasks, provides a solid foundation for feature extraction in medical image classification. To adapt to the osteoporosis three-class classification task (normal bone mass, osteoporosis, and osteopenia) of this invention, the last fully connected layer of the VGG16_bn model was modified to output parameters for the three classes. Before inputting the enhanced images into the VGG16_bn model, a resize operation is performed to uniformly compress them to 224×224 pixels to meet the input size requirements of the VGG16_bn model. In particular, considering the significant class imbalance problem in the training set (as shown in Table 1), this embodiment innovatively introduces the Focal Loss function during the training process of the VGG16_bn model. The Focal Loss function dynamically adjusts sample loss weights, enabling the model to more effectively focus on and learn those difficult-to-classify minority class samples. This better handles class imbalance during training, ensuring the model's ability to recognize all classes is balanced and accurate. The use of pre-trained models significantly reduces training time and substantially improves initial performance and final accuracy. Figure 4 As shown, the entire training process involves loading pre-processed training data, fine-tuning the VGG16_bn model, and continuously updating the network parameters through backpropagation, so that the prediction results of the VGG16_bn model approximate the true labels.

[0043] During the model validation and evaluation phase, the validation set data, which underwent standardized preprocessing and two-stage augmentation, was loaded into the trained classification model. The classification model used a classifier to derive the probability distribution of each image belonging to each category, and then used five weighted sums of augmentations to select the category with the highest probability as the final classification result. The validation set used and the accuracy of classification for different categories are shown in Table 2.

[0044] Table 2 shows the accuracy of classification for different categories on the validation set. This embodiment achieved a classification accuracy of 98% on the validation set, demonstrating the superior performance of the method proposed in this invention. These results fully demonstrate that the method proposed in this invention can significantly improve the classification accuracy of the model. Upon final deployment, this invention will provide doctors with efficient and reliable auxiliary diagnosis, effectively reducing the probability of misdiagnosis, thereby greatly improving the work efficiency of doctors and hospitals.

[0045] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0046] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0047] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A deep learning-based method for classifying wrist joint bones, characterized in that, The method includes the following steps: The original DICOM format medical images are standardized, and a two-stage data augmentation operation is performed on the standardized medical images to construct a medical image dataset divided into training and validation sets based on wrist joint bone types. The two-stage data augmentation includes: In the first stage, we will perform integrated adaptive local contrast enhancement, edge feature enhancement, and unified preset enhancement for pseudo-color rendering. In the second stage, real-time random data augmentation operations, including random horizontal and vertical flips, random rotations within the range of [-20°, +20°], and random cropping and scaling, are dynamically injected into each epoch of the VGG16_bn model training. The VGG16_bn model, pre-trained on the ImageNet large-scale dataset, was loaded as the backbone network. The VGG16_bn model was trained and optimized using the training set to obtain a classification model for intelligent classification of wrist joint bones in medical images. The classification model's performance was validated and evaluated using a validation set to obtain its classification accuracy, thereby quantifying the model's effectiveness and potential in wrist joint bone classification.

2. The wrist joint bone classification method according to claim 1, characterized in that, The standardization process for the original DICOM format medical images includes: First, the pydicom tool is used to accurately parse the acquired DICOM format medical image files, and to fully extract their pixel data and associated metadata. Secondly, window width and window level normalization operations are performed on the extracted pixel data to linearly map the original high dynamic range pixel values ​​to the standard 0-255 range. The processing logic for MultiValue pixels is specially optimized to ensure that each pixel is calculated with a single, accurate value. Finally, the processed medical images are scaled to a target resolution of 1024×1024 according to the principle of proportional scaling, and saved as .png format while maintaining the original image proportions and avoiding distortion.

3. The wrist joint bone classification method according to claim 1, characterized in that, The constructed medical image dataset includes: First, the standardized medical image dataset is divided according to the image category, with the ratio of training set to validation set set being 85%:15%, and ensuring that each image category is evenly distributed in both the training set and the validation set. Secondly, the first phase of unified preset enhancement is implemented on the divided medical image dataset to enrich image features from a global perspective and make up for uneven lighting and low contrast problems during the image acquisition process. Finally, during each epoch of training the VGG16_bn model, a second phase of real-time random data augmentation is dynamically applied to the medical images in the training set.

4. The wrist joint bone classification method according to claim 3, characterized in that, The image categories include normal bone mass, osteoporosis, and reduced bone mass.

5. The wrist joint bone classification method according to claim 1, characterized in that, The training and optimization of the VGG16_bn model includes: Model initialization and input adaptation: The VGG16_bn model, pre-trained on the ImageNet large-scale dataset, is loaded as the backbone network for deep feature extraction. At the same time, the medical images that have undergone two-stage data augmentation are uniformly adjusted to the standard input size of 224×224 required by the VGG16_bn model. Fine-tuning training and loss function optimization: The VGG16_bn model is fine-tuned using the training set. During training, the Focal Loss loss function is introduced to evaluate the difference between the model output and the true label. The SGD optimizer is used to dynamically adjust the network weights. The VGG16_bn model predicts the probability of each category through forward inference of a neural network and continuously optimizes the network parameters through backpropagation.

6. The wrist joint bone classification method according to claim 5, characterized in that, The top-level classifier of the VGG16_bn model was modified to be a fully connected layer that outputs three categories.

7. The wrist joint bone classification method according to claim 6, characterized in that, The three categories are normal bone mass, osteoporosis, and osteopenia.

8. The wrist joint bone classification method according to claim 1, characterized in that, The verification and evaluation of the classification model's performance includes: Validation set inference and probability output: The validation set is loaded into the trained classification model. The classification model performs forward inference on each medical image in the validation set and transforms the original score vector into a normalized probability distribution through the SoftMax function connected to the output layer. The class with the highest probability value is selected as the predicted classification result for the image. Accuracy statistics and comprehensive evaluation: By comparing the predicted classification results of the classification model with the real labels of medical images in the validation set one by one, the number of correctly classified samples is recorded, and the classification accuracy of the classification model is calculated.

Images